Terafab Vertical Integration: Who Controls the Stack From Silicon to Orbit?

Terafab is the planned Grimes County semiconductor complex through which SpaceX and Tesla want to bring chip design, logic, memory, packaging and testing into one coordinated system. SpaceX estimates the initial phase at $16.8 billion, more than 100 million square feet and at least 3,000 jobs. The strategic end point is bigger: manufacture compute, launch it with Starship, then run AI workloads through SpaceX and xAI infrastructure.

That is vertical integration from silicon to orbit. It could shorten feedback loops and protect supply. It also concentrates capital allocation, intellectual property, infrastructure access and platform power across companies controlled by the same executive. The key question is therefore not only whether Terafab can be built. It is who owns each layer, who can buy access, and which safeguards keep the stack contestable.

Terafab facts in 60 seconds

SpaceX's 7 August 2026 Terafab update confirms Grimes County and describes the first phase as an estimate, not completed capacity. That distinction matters when headlines turn a plan into an operating fact.

Texas JETI applications filed by Terafab AI, LLC describe a vertically integrated campus with up to four phases, sub-2 nm class technology and a planning range of $55 billion to $119 billion. The public Texas Comptroller application is evidence of the proposed scope and incentive request. It is not evidence that the upper figure has been spent or contractually committed.

What does “silicon to orbit” actually include?

The phrase is a useful map, not a literal claim that one company mines sand and owns every supplier. Terafab is designed to close several adjacent layers that currently sit across different vendors:

1. Workload and model demand. xAI, Tesla autonomy, Optimus and third-party compute customers create demand for training and inference.
2. Chip architecture. The program targets edge and inference chips for Tesla hardware plus high-power devices designed for space compute.
3. Process and masks. The stated model includes photomask generation and semiconductor process work, not only chip design.
4. Logic, memory and packaging. Co-locating these stages is intended to reduce handoffs and make hardware iteration faster.
5. Systems and satellites. Chips become compute payloads, servers, robots, vehicles and orbital AI satellites.
6. Launch and operation. SpaceX supplies the launch system, satellite network and operating platform that can deploy and sell the resulting compute.

That chain is why Terafab belongs beside Starmind, not merely beside other chip factories. The fab is an upstream capacity strategy for a downstream compute platform.

Why can extreme vertical integration make economic sense?

Semiconductor manufacturing is difficult to enter because early yields are uncertain and a new foundry must persuade customers to risk designs on an unproven process. Terafab starts with captive demand. SpaceX and Tesla can consume the output themselves if the chips meet their requirements. That does not remove yield risk, but it reduces demand risk.

• Supply assurance: reserved internal capacity can protect programs whose growth would otherwise depend on a small set of external fabs and packaging suppliers.
• Faster feedback: model, chip, packaging, system and deployment teams can trade data without long commercial handoffs.
• Hardware-workload co-design: a space chip, a robot inference chip and a general data-center GPU do not need the same power, cooling or reliability profile.
• Margin capture: successful internal production can remove supplier margins, although low yields and underused capacity can erase that advantage.
• Scheduling control: the owner can prioritize the chips that unblock its own product roadmap.

The commercial lesson is narrower than “build everything.” Own the constraint only when it is durable, differentiating and expensive enough that supplier dependence changes your strategy. For most product companies, portability and strong vendor contracts beat ownership of the physical stack.

Who owns Terafab today?

The most accurate answer is that public disclosure describes a collaboration, not a simple fully disclosed joint venture with settled economics. SpaceX acquired xAI effective 2 February 2026, so xAI is not an independent counterweight. Tesla remains a separate public company. Intel is a strategic partner.

SpaceX's registration statement and risk disclosures say that SpaceX and Tesla have a general framework agreement. Specific projects still require separate negotiation and board approval, and neither Tesla nor Intel is obligated to remain. The filing also says each party retains pre-existing and independently developed intellectual property and that SpaceX expects to keep sourcing a significant share of compute hardware from third parties.

Those boundaries make five questions commercially material:

• Capital: which company funds overruns, equipment and later phases?
• Allocation: who gets scarce good wafers when Tesla, SpaceX and external buyers all need supply?
• Intellectual property: who owns process improvements and chips co-developed across company teams?
• Transfer pricing: how are wafers, power, launches and compute priced between related parties?
• Exit rights: what happens to capacity and know-how if a partner leaves?

This is not a claim that the structure is improper. It is a list of facts buyers, boards and investors need before “joint venture” becomes a substitute for a contract map.

Intel 14A adds capability, but not contractual certainty

Intel joined the program in April. Its Q1 2026 earnings comments confirm the partnership with SpaceX, xAI and Tesla, while discussing active external evaluation of Intel 14A. Intel brings process development, fabrication and advanced packaging experience that a new entrant cannot recreate quickly.

Public statements have connected Terafab with Intel 14A, but the disclosed framework still leaves the final division of labor, process licensing, equipment ownership, volume and schedule open. “Intel joined” should not be translated into “Intel guarantees one million wafers per month.” The partnership improves plausibility. It does not cancel execution risk.

What the Nvidia Starmind partnership changes

Two days before the Grimes County confirmation, SpaceX said Starmind's first AI satellite payload would use Nvidia Vera Rubin systems. The official Starmind overview places localized compute, solar power, radiators and laser connectivity in orbit.

That makes Nvidia and Terafab complements on different time horizons. Nvidia supplies a proven near-term compute platform. Terafab is an attempt to create additional, workload-specific supply and bargaining power later. SpaceX's own filing expects meaningful third-party hardware purchases to continue.

The stronger interpretation is therefore not “Musk replaces Nvidia.” It is “Musk wants an internal option at the chip layer while buying the best available external systems.” Good technology strategy often looks like this: preserve a credible supplier path while funding a differentiated internal path behind measurable gates.

The competition question nobody should postpone

A vertically integrated stack can be both efficient and open. SpaceX says competitors can buy launches, connectivity and compute. The governance test is whether access remains commercially realistic when the same group also competes downstream.

The risk is not vertical integration by itself. The risk is a closed control point whose rules are unclear, whose switching costs compound across layers and whose customers cannot verify equal treatment.

A build, buy or partner test for your own AI stack

Terafab is an extreme case, but its decision logic applies to smaller systems. Use five gates before moving any AI infrastructure layer in-house:

1. Name the strategic constraint. Is the blocker cost, latency, supply, data control, compliance or product differentiation?
2. Prove captive demand. Measure the stable workload that will use the asset, not the total addressable market in a slide.
3. Price the feedback loop. Quantify how much faster product learning becomes when adjacent teams share telemetry and release cycles.
4. Keep an external option. Define interfaces, data export, fallback capacity and exit terms before internal integration makes them expensive.
5. Stage capital behind evidence. Fund the next layer only after utilization, quality and operating cost pass a written threshold.

For the chip-level version of specialization, our Taalas HC1 hardwired LLM ASIC review shows when speed can justify model lock-in. For a broader sourcing decision, use the custom software versus off-the-shelf framework. Both separate a differentiated constraint from a fashionable desire to own more.

Wavect applied the same boundary-first thinking while building Prompt.ID's production AI platform: architecture needs clear ownership, measurable quality and recoverable supplier choices. If your roadmap spans models, data, infrastructure and product workflows, an AI architecture and product engineering engagement can turn those layers into a staged decision rather than one irreversible bet.